Sharing the Smooth Background Across a Jump: Enriched Fourier Extensions with Truncated SVD

Fourier methods are fast and very accurate for smooth periodic signals. They break down whenever the signal has a jump. They also break down at the ends of a non-periodic interval, because periodic repetition creates an artificial jump there. In both cases the result is the Gibbs phenomenon: ringing near the jump that overshoots by about 9% of the jump height and never goes away as more modes are added (The World of Fourier Series). Fourier extension removes the boundary version of this problem. It fits the data on $[-1,1]$ with Fourier modes that are periodic on a larger interval $[-T,T]$, so the fitted function can bend smoothly through an unused buffer zone. The resulting linear system is exponentially ill-conditioned. Truncated singular value decomposition (TSVD), which solves the system while discarding directions with tiny singular values, stabilizes it. This works because those discarded directions live almost entirely in the buffer zone (On the Numerical Stability of Fourier Extensions). Interior jumps, such as shocks in a fluid, still defeat the method. Current remedies either cut the domain at detected edges and fit each piece separately (Reconstruction of Piecewise Smooth Functions Based on Fourier Extension), or add viscosity chosen by a neural network (FC-based shock-dynamics solver with neural-network localized artificial-viscosity assignment).

We propose a third option: do not cut the domain. The idea is to fit one global Fourier extension to all the samples, add a small block of "jump atoms" (step functions times low-degree polynomials, anchored at an unknown jump location $\xi$), and solve the combined system with a single TSVD. The jump location is found by a one-dimensional search that minimizes the TSVD residual. A smoothness penalty is applied only to the Fourier block, never to the jump atoms. The bet is that many discontinuous signals of interest, such as a shock riding on a smooth flow, consist of a smooth background shared by both sides plus a jump correction of low complexity. Cutting the domain throws that sharing away, and the loss hurts most when one side of the cut is short or the data are noisy.

Core Claim

Write the reconstruction as

$$ f(x) \approx \sum_{|n|\le N} a_n \phi_n(x) \;+\; \sum_{k=1}^{K} H(x-\xi_k)\sum_{j=0}^{q} b_{kj}\, p_j\!\left(\frac{x-\xi_k}{1-\xi_k}\right), $$

where $\phi_n(x)=e^{in\pi x/T}/\sqrt{2T}$ are the extension modes, $H$ is the unit step, $\xi_k$ are the jump locations, $p_j$ are Legendre polynomials rescaled to the right-hand segment, and $a_n, b_{kj}$ are coefficients. For fixed locations the model is linear. Let $A_\xi$ be its $M\times(2N+1+K(q+1))$ sample matrix and $A_\xi^{\dagger,\tau}$ its weighted TSVD pseudo-inverse with threshold $\tau$. The locations are then estimated by

$$ \hat\xi = \arg\min_{\xi}\ \big\| (I - A_\xi A_\xi^{\dagger,\tau})\, f_{\text{samples}} \big\|_2 . $$

This is variable projection, the standard way to solve separable nonlinear least squares by optimizing out the linear coefficients. We make three claims.

First, the enriched system keeps the property that makes Fourier extension work: its small singular vectors stay concentrated in the buffer zone. TSVD at a threshold of $10^{-10}$ to $10^{-14}$ should therefore remain stable. We expect the jump atoms to sit at $O(1)$ singular values, because no smooth periodic function can mimic a step on $[-1,1]$.

Second, the residual objective has a sharp minimum at the true $\xi$. That gives sub-grid-cell location accuracy and jump heights directly. The coefficient-norm spike that A Local Fourier Extension Method for Function Approximation uses as a binary "smoke detector" becomes the initializer of a continuous estimate.

Third, applying the frequency-weighted penalty of A New Approach for Fourier Extension Based on Weighted Generalized Inverse to the Fourier block alone resolves the conflict the research file flags. That penalty suppresses buffer ringing and stabilizes derivatives, but it would smear a physical jump. Here the jump is carried by unpenalized atoms, so penalizing high frequencies costs nothing at the discontinuity.

The method is nonlinear only in $\xi$. This matters because A Stability Barrier for Reconstructions from Fourier Samples proves that stable linear schemes can reach only root-exponential convergence when resolving Gibbs, and Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities proves a power-law floor for linear reconstructions of moving shocks. Neither result binds a method that adapts its basis to the data. Both are why the location search cannot be skipped.

Relation to Prior Work

Two papers motivate the gap and supply the core machinery. Parameter selection and numerical approximation properties of Fourier extensions from fixed data established TSVD-regularized extension from equispaced data, the $T=2$ convention, and the precision-resolution trade-off controlled by $\tau$. It treats only smooth functions. We reuse its solver and parameter choices. On the Numerical Stability of Fourier Extensions supplies the buffer-localization argument we must re-verify for the enriched system.

The closest competitors handle interior jumps by partitioning. Reconstruction of Piecewise Smooth Functions Based on Fourier Extension detects edges with mollifiers, splits the domain, and fits local extensions, reaching $10^{-14}$ to $10^{-15}$ errors. A Local Fourier Extension Method for Function Approximation partitions uniformly, reuses one TSVD across pieces, and flags singular pieces by coefficient-norm spikes. Both fit each side of a jump only from that side's samples. Neither returns a sub-cell jump location or height. Our approach keeps one global smooth fit.

The closest structural relative is Hybrid Surrogate Models. It also adds Heaviside terms to a smooth expansion, but it uses Chebyshev polynomials and gradient-based training. Chebyshev fits are unstable on equispaced samples (the Runge phenomenon), and that work never considers the ill-conditioned frame. Batenkov's decimated Eckhoff and Prony methods (Complete algebraic reconstruction of piecewise-smooth functions from Fourier data, Prony systems via decimation and homotopy continuation) recover jumps algebraically with optimal rates, but they need exact Fourier coefficients of a periodic function and the jump count in advance. Our setting is point samples on a non-periodic interval.

Research Program

The first stage is analysis of the enriched matrix in 1D. We compute its singular spectrum as $N$, the oversampling ratio $\eta=M/(2N+1)$, the polynomial degree $q$, and the distance of $\xi$ to the boundary vary. We measure what fraction of each small singular vector's energy lies in the buffer zone, and test whether interlacing with the plain extension spectrum holds. If small singular vectors leak into $[-1,1]$ when $\xi$ approaches $\pm1$, that is a real failure mode. We would then add a boundary-specific extension, in the spirit of the boundary-focused stage of Li, Zhao, and Jia.

The second stage studies the location objective. We characterize the width of its basin, how the minimum's sharpness scales with jump height and noise, and how to handle the number of jumps. Our plan is to add jumps greedily while the residual drops by more than a noise-calibrated factor, which removes the a priori jump count required by Prony methods.

The third stage is convergence. When the jump correction is a polynomial of degree at most $q$, the remainder is smooth and should converge at the geometric rate of a smooth Fourier extension. When it is merely analytic, we expect convergence that is exponential in $q$. Stability on equispaced points caps $q$ near $\sqrt{M}$, which mirrors the equispaced polynomial constraint and predicts root-exponential behavior in that regime. We will test both predictions.

The fourth stage takes the method to shock data: 1D Burgers snapshots, then tracking $\hat\xi$ across time.

Baselines

The naive reference is a plain TSVD Fourier extension at $T=2$ and $\tau=10^{-14}$, following Adcock and Ruan. It is run on identical samples and should show the full interior overshoot. Any method that does not substantially remove it fails.

The strongest existing method is the edge-detect-and-split extension of Li, Zhao, and Jia. It reports $10^{-14}$ to $10^{-15}$ reconstruction error on its synthetic piecewise-smooth suite, and that is the bar on clean, well-separated jumps. We will re-implement it with the same $T$, $\tau$, and per-piece oversampling, and give it oracle jump locations as well as its own detector. That separates detection error from fitting error. We also run the local TSVD extension of Zhao and Wang with adaptive subdivision. Beating these baselines means matching their error to within one order of magnitude on well-separated jumps, while winning by at least two orders in L∞ error on the short-piece, close-jump, and noisy regimes. We also require jump-location error below a tenth of the grid spacing, which split methods cannot report at all.

The HSM-style competitor follows Chirra et al.. That paper reports an $l_1$ error of $7.33\times10^{-5}$ on its shock reconstruction task, against $4.92\times10^{-3}$ for the best plain network. We will run it on equispaced samples, both as published and as a Chebyshev-plus-Heaviside least-squares fit under our TSVD and location search. The second version isolates the contribution of the Fourier extension frame.

For learned operators on Burgers, Lanthaler et al. report a relative median $L^1$ error of 1.57% for a Fourier Neural Operator and 28.5% for DeepONet. Those numbers measure operator learning rather than snapshot reconstruction, so they serve only as context.

Ablations isolate the key ingredients. Removing the global Fourier fit across the jump turns our method into a split method expressed in one frame, which tests whether the shared background matters. Fixing $\xi$ at cell midpoints tests the continuous search. Placing the frequency weight on no block, on the Fourier block only, and on all blocks tests the resolution of the smoothness-penalty conflict. Swapping Tikhonov regularization for TSVD tests the choice of regularizer. All methods receive the same samples and the same tuning grid over $N$, $\tau$, and $q$, selected on held-out functions.

Why not just…?

Why not just split the domain at detected edges, as Li, Zhao, and Jia do? This is the serious threat. When the two sides of a jump are unrelated and both pieces are long, splitting should match us. We will report that regime honestly as a tie or a loss. Splitting loses information when a piece is short. The resolution analysis of On the resolution power of Fourier extensions for oscillatory functions puts the required points per wavelength at roughly $T\eta$, so a piece with few samples cannot resolve background oscillation that the global fit sees in full. Splitting is also vulnerable to noisy edge detection. The experiment that settles the question sweeps the distance from the jump to the boundary and the spacing between jumps, at several noise levels, against oracle-location splitting. If oracle splitting matches us everywhere, the shared-background idea is wrong and the remaining contribution shrinks to sub-cell location.

Why not just filter or add artificial viscosity? Filtering trades the overshoot for smearing across several cells, and global filters drop accuracy to low algebraic order. Leibovici et al. needed a trained detector precisely to keep viscosity out of smooth regions. We include exponential-filtered extension as a baseline and report error as a function of distance from the jump, which shows the smearing directly.

Why not just use the frequency-weighted TSVD of Zhao alone? That penalty is designed for smooth targets. Applying it to a function with a jump suppresses exactly the slowly decaying high frequencies the jump needs, which is the "Sobolev penalty paradox" in the research file. The weight-placement ablation tests this directly. It should show that weighting everything blurs the jump, while weighting only the Fourier block keeps derivative accuracy without blurring.

Why not Chebyshev plus Heaviside, which has no buffer zone? On equispaced data, high-degree polynomial fits are unstable, while Fourier extension with linear oversampling is stable (On the Numerical Stability of Fourier Extensions). Polynomial grids also force tiny time steps in PDE solvers. The HSM-style ablation shows whether this matters at the sample sizes we test.

Proposed Experiments

Static tests use the nine-function suite of Adcock and Ruan and the four functions of Zhao ($e^t$, $\mathrm{erf}(10t)$, $1/(1.2-t^2)$, $\cos(10/(1+25t^2))$). We add one to three jumps of controlled height, derivative order, and position, and add Gaussian noise at $10^{-8}$, $10^{-4}$, and $10^{-2}$. Metrics are L∞ error at distance greater than one grid cell from any jump, $L^1$ error overall, residual overshoot as a fraction of jump height, jump-location error in grid cells, jump-height error, first-derivative error away from jumps, and wall-clock time.

Shock tests use 1D Burgers snapshots from PDEBench, subsampled to non-periodic windows so that boundary and interior Gibbs effects co-occur. We report location error against the high-resolution reference and the consistency of $\hat\xi(t)$ across snapshots. We found no reported result on Lacuna for any extension-based reconstruction of these snapshots, so all comparisons there are run by us.

Four checks decide the outcome. The first is whether small singular vectors of the enriched matrix keep more than 99% of their energy in the buffer zone. The second is whether location error falls below a tenth of a cell at moderate noise. The third is whether we beat oracle splitting by two orders of magnitude when one piece contains fewer than about $4T\eta$ samples. The fourth is whether derivative error with the Fourier-block-only weight matches Zhao's smooth-case behavior while the jump stays sharp.